A lorry is supposed to arrive between 1pm and 3pm, It’s now 10 past 3, and nobody knows if it’s 2 minutes away or 5 hours away. Its slot is still in the system, there’s another vehicle sat at the gate, and the warehouse team are working to a plan that stopped being true.
Nobody reading this will find that unusual, it’s common everywhere. Spotting it isn’t really the problem, the problem is getting everyone onto the same plan before a small delay turns into a queue, and later into a heated conversation about who knew what.
We started Finmile as a delivery company, doing about 10 parcels a day, and grew it to 55,000 a day. Along the way we used a lot of software, and the same thing kept happening. We’d ask a vendor for one small change and be told, sure, no problem, 6 months and £50,000. So we built our own, and even once we had good tools, we kept finding jobs where someone still had to pick up the phone, chase an answer or get 2 teams talking to each other.
That’s the gap I’ve become wary of: the distance between knowing something’s wrong and actually sorting it. A dashboard will tell you the plan has broken, it won’t ring the haulier. That’s where AI can genuinely earn its place in the yard, as long as we’re honest about what we’re asking it to do.
I’m not knocking visibility, arrival estimates and dock availability are a huge improvement on guesswork and I wouldn’t want to go back. But think about the train stations, before the countdown screens went up, passengers on the platform had no idea if the train was 2 minutes delayed or 20, and they were furious. Then the boards started saying, sorry, this train is delayed by 4 minutes, and everyone just sat down and got their phones out. The delay didn’t change, what changed was that the unknown went away.
Yards are the same, except an alert on a screen still leaves one person doing everything, they phone the haulier, check with the warehouse, find another slot, explain it to the driver, update the booking, and then, quite often, do it all again because something else has moved. Meanwhile the warehouse is waiting and you’re paying people overtime to stand around, and those knockon costs from even 30 minutes late are the bit people tend to underestimate.
Usually the software exists, what’s missing is the joining up. The booking lives in the yard system, the journey lives in the transport system, the receiving plan sits with the warehouse, and none of them on its own tells you what’s really going on. Plenty of the drivers involved are independent contractors, too, which means you can’t simply track them and call it done. So if AI just produces another summary for someone to read before making the same calls, I’d argue we haven’t saved anyone any work at all.
Say a vehicle looks likely to miss its slot, first you need to know what’s actually true. Is the ETA current? Has the driver confirmed it, or is it a prediction? Those are very different things and the system has to keep them apart. Then the practical questions. Is there another slot? Can receiving cope with the load at that time, and is there anything about the goods or the unloading order that rules it out?
AI can chase those answers through approved channels, whether that’s a call, a message to the driver or a note in Teams or Slack, and put an alternative together. Where the site’s rules allow, it can confirm the change and update the booking. Where they don’t, it hands the decision to whoever owns it. And it does all of that calmly and the same way every time, which is more than any of us can promise at the end of a long shift when we’re annoyed.
It also has to keep going after the message is sent, did the haulier accept? Does the warehouse know? Has the lorry actually turned up? If nobody checks, you’ve just got a neater record of a problem nobody solved.
Asking a driver for an update is one thing, changing something that affects safety, cost or someone else’s delivery is quite another, and I don’t think that should ever be hidden inside an instruction as woolly as “optimise the yard”. Operators should write down what the system can do on its own, what needs sign off, and when it has to stop. Status checks could be automatic. Moving a booking might need a supervisor. Anything involving a charge or a safety concern probably stays with a person, full stop.
A booking change is not permission to move a vehicle, site traffic rules and pedestrian separation still apply, and a language model doesn’t replace the people responsible for keeping a yard safe. Gaps matter as much as rules. If capacity can’t be confirmed, don’t promise the slot. If 2 records disagree, say so. If nobody replies, a named person picks it up, because silence isn’t agreement. I’d always take a system that asks a slightly annoying question over one that confidently makes a mess for the night team.
“Human review required” on its own is close to useless. When something lands on a supervisor’s desk they need to see what happened, what’s confirmed, who’s been contacted and when a decision is due, so they aren’t redoing the investigation. You also want a trail afterwards: who approved the new slot, which systems changed, whether anyone acknowledged it, and whether it can still be undone.
I wouldn’t try to make a whole yard autonomous at once, the way we work with customers is to spend a lot of time listening, pick one workflow that keeps coming back, agree a hypothesis for what fixing it is worth, and prove it before touching anything else. If it doesn’t pay for itself, there’s no point doing it. We let the system suggest before it acts, compare its suggestions with what the team actually did, spend time on the awkward cases, and give it more room only when the evidence is there.
The measure I care about is how long it takes to go from spotting a problem to a confirmed fix. After that, count the repeat calls, the pointless escalations, the wrong changes and the problems that come back after being closed. And check waiting times improve without something else quietly getting worse. Messages sent and alerts closed don’t count, what counts is a plan the people involved can actually carry out, and a record of what happened.
Yards will always have exceptions, what we can do is cut the chasing, and leave judgement with the people who ought to have it. Visibility tells you the plan’s broken. Useful AI should help you put a new one together, that’s exactly what Finmile Autonomy was built for.
hello@finmile.co